subagentprompts

.com prompt engineering
the prompt-engineering primitive

One Sonnet main model. Fable 5, twice, for a reason.

This Cowork session runs on Claude Sonnet 5 as its main model — that's the whole session, not a claim this site makes about itself. Rather than invent a hypothetical multi-model workflow, this site documents the real one: Anthropic's advisor tool explicitly accepts a Sonnet main model paired with a Fable 5 advisor, "Fable 5 guidance at decision points without running Fable 5 throughout." This session used Fable 5 twice, in two different real roles: once as the primary builder of subagentcode.com (a full Agent-tool dispatch, model:"fable"), and once as an advisor reviewing this very site's draft before deploy — the second role is the one the real advisor feature actually describes.

Start here: the live board of prompting techniques →

Grounded, not generated

Every row in this catalog cites a real doc — either already mirrored in this repo's docs/ tree, or fetched live from platform.claude.com this session. Nothing paraphrased from memory.

Model-specific first

Matches the real structure of Anthropic's own prompting-best-practices reference: model-specific guidance, then general techniques, then migration notes.

Real gotchas included

Including the ones that are easy to miss — like prefill not being supported on Fable 5, Mythos 5, Opus 4.6+, or Sonnet 4.6.

Live board

All rows are seeded verbatim from the source docs into this Worker's own D1 database (prompt_techniques). Browse the board → or hit the JSON API.